Visualizing Data Transformations Through the Pipeline
This dashboard visualizes how data transforms as it flows through the GNN pipeline:
Raw Time Series Data → Find Continuous Segments → Create Windows → Compute Adjacency Weights → Batch Data → Model Forward Pass
Starting point: time series data with missing values (shown as gaps).
Identifying continuous segments in the time series data without significant gaps.
Creating fixed-size windows from continuous segments, standardizing values, and tracking missing data with masks.
Converting distance matrix to weighted adjacency matrix using Gaussian kernel.
Organizing data into batches for model training with shape transformations.
Visualizing the STGNN model architecture that processes the tensor data.
Summary of how tensor shapes change through the pipeline:
Raw Time Series: Series[time] → shape: (n_timestamps,)
Windowed Data: X_by_sensor[node_id] → shape: (n_windows, window_size)
Batched Data: x → shape: (batch_size, num_nodes, seq_len, features)
Model Output: predictions → shape: (batch_size, num_nodes, horizon, features)
Raw Adjacency Matrix → shape: (num_nodes, num_nodes)
Weighted Adjacency Matrix → shape: (num_nodes, num_nodes)